Claim Assistant
AI-powered automation that streamlines insurance claim intake with LLM-based PDF processing, policy matching, and coverage analysis
Highlights
A large insurer processed ~107K claims per year across phone, email, fax, and portal, with 30 staff manually re-keying data into ASC, limited to 12h/day on weekdays.
Build an automation solution to extract claim data from PDFs, faxes, and handwritten forms, score per-field confidence, and map fields to ASC for 24/7 processing.
Designed a two-stage pipeline pairing Azure Document Intelligence OCR with GPT-5 mapping and validation, plus a review UI with bounding-box highlights and confidence-gated export.
Cut processing to ~1 minute per form with 24/7 pre-processing that eliminates backlog, focusing staff on low-confidence exceptions and reducing data-entry errors.
Core Team
Overview
Claim Assistant is an AI-powered solution that automates insurance claim intake and processing, transforming a slow, manual workflow into a fast, scalable, and accurate one. It converts filled insurance claim forms, including scanned and handwritten documents, into structured, validated data ready for downstream systems.
The platform pairs specialized document intelligence with LLM-based reasoning. Azure Document Intelligence extracts key-value pairs, layout, and per-field confidence, while OpenAI GPT-5 maps the results to the target schema, resolves field aliases, and generates a coverage summary. A review interface with side-by-side PDF previews and bounding-box highlights lets adjusters verify, edit, and approve fields, with a confidence-driven queue that focuses attention only on the exceptions that need a human.
Data
The solution targets a high-volume claims operation handling roughly 107,000 claims per year (~9,000 per month) arriving through multiple channels:
- •ASC System (75%): ~6.8K claims/month from phone (37%), email (45%), and fax (18%) channels.
- •External Portal (25%): ~2.5K claims/month submitted directly by insureds through partner portals.
Inputs span PDFs, faxes, and emails, both digital and handwritten, across claim forms from eight US states (Florida, New Hampshire, Minnesota, Iowa, Kansas, New York, Ohio, and Wisconsin). The pipeline is form-agnostic, handling arbitrary layouts without per-form training, and flags any field extracted below 80% confidence for human review.
Methods
Claim Assistant uses a two-stage pipeline that combines specialized document intelligence with LLM-based mapping and validation:
- Stage 1 (Document Intelligence): Azure Document Intelligence (Form Recognizer v3.x) extracts key-value pairs, bounding boxes, layout structure, and per-field confidence scores from each document.
- Stage 2 (LLM Mapping & Validation):
- •Field Mapping: GPT-5 aligns extracted values to target schema fields while preserving evidence links to the source document.
- •Alias Resolution: Normalizes inconsistent labels and synonyms across form variants into a single canonical schema.
- •Policy Matching: Maps extracted identifiers to policy records using weighted Levenshtein similarity and date validation.
- •Summary Generation: Produces a coverage analysis (covered / not covered / uncertain) with transparent reasoning and confidence metrics.
- Review Workflow: A side-by-side review UI highlights each field on the source PDF, supports inline editing and approval, surfaces a dedicated low-confidence queue (<80%), and gates export to ASC until all fields are approved.
Results
By shifting intake from full manual data entry to AI-driven pre-processing, the solution delivers measurable operational gains:
- •Processing Speed: ~1 minute per form (parallelizable), down from longer, fully manual data entry.
- •Coverage Hours: 24/7 automated pre-processing replaces 12h/day weekday-only coverage, eliminating overnight and weekend backlog.
- •Staff Productivity: The confidence-driven queue lets staff focus only on exceptions and low-confidence cases instead of re-keying every claim.
- •Error Rate: Reduced through automated validation and visual bounding-box verification.
Being form-agnostic, the pipeline generalizes across layouts and input quality, from clean digital PDFs to noisy handwritten submissions, without per-form training:
Conclusion
Claim Assistant shows how pairing document intelligence with LLM reasoning can turn a manual, bottlenecked claims operation into a scalable, 24/7 pipeline. Its hybrid approach (Azure for accurate extraction, GPT-5 for mapping and validation) combined with a confidence-driven, evidence-linked review workflow, keeps humans in the loop exactly where it matters while automating the rest.
The architecture is ready for enterprise integration. Roadmap items include a document classification pre-filter to route incoming messages (FNOL, billing, misrouted), direct email-body parsing, and full ASC export integration for end-to-end straight-through processing.